v0.12

• As always, more models are supported.
• Multimodal: vision & TTS.
• Windows: prebuilt binary with Vulkan (1.4.304.1). Use -ngl all to run whole model on default GPU.